Predição do desempenho fisiológico e de CO2 em soja: uma abordagem utilizando sensoriamento hiperespectral e aprendizagem de máquina

Abstract

Hunger and climate change plague the planet, and actions aimed at mitigating them have been the focus of government entities and research centers. Techniques that combine these two factors are fundamental for life and nutrition. Environmental sciences associated with soybean breeding, precision agriculture and machine learning (ML) are an alternative to this hardship. This work aims to predict physiological variables, in situ emission of soil carbon dioxide (FCO2) and carbon fixation in leaf tissue in soybean genotypes through hyperspectral variables and ML, as well as identifying the best algorithms. A randomized block design with four replications was used. The plots consisted of five rows of five meters with a spacing of 0.45 m between rows. The assessments were: net photosynthesis; stomatal conductance; internal CO2 concentration; perspiration; instantaneous water use efficiency; instantaneous carboxylation efficiency; plant spectral analysis; carbon dioxide flow, in situ soil temperature and humidity, and carbon fixation in leaf tissue. The ML models used were: Artificial neural networks; REPTree decision tree; Decision tree (M5P); Random Forest (RF); Support vector machine; and zero R. The parameters used were: Pearson correlation (r), mean absolute error and square root of the mean error. It is possible to predict net photosynthesis with r above 0.75, which is excellent, and the other physiological variables permeate average results. Carbon flow and carbon fixation showed unsatisfactory results, which requires further studies on these predictions. The best ML techniques for this approach are: RF and M5P.